Clamping SAE batch-effect features corrects single-cell batch integration

measured in 1 paper

Pedrocchi et al. train BatchTopK SAEs on residual-stream representations of single-cell foundation models scGPT and scFoundation across several datasets [pedrocchi-etal-2025-scfm-saes] Batch/technical-effect feature directions are identified via mutual information between feature activation and batch label [pedrocchi-etal-2025-scfm-saes] Clamping the top batch features (top-20 pretrained, top-50 fine-tuned) and re-decoding improves scIB batch correction without substantial loss of biological conservation [pedrocchi-etal-2025-scfm-saes] The intervention shows a dose-response against random-feature-ablation, peaking at dataset-specific thresholds (25 Pancreas, 30 Lung, 60 Immune) before plateauing or degrading [pedrocchi-etal-2025-scfm-saes]

Structure

Linear Direction

Context

BatchTopK SAEs trained on scGPT and scFoundation residual streams, mutual-information-based identification of batch/technical-effect feature directions, feature clamping (causal suppression) and re-decoding through the frozen model, dose-response validated against a random-feature-ablation control (25/30/60-feature thresholds), standard scIB batch-correction and bio-conservation benchmark scores

Confirmed in models

scGPT (whole-human pretrained checkpoint) · scGPT, scFoundation (pretrained) · scFoundation

Method

BatchTopK Sparse Autoencoders, Causal interventions (steering)

Papers

Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models — Pedrocchi, Flavia, Barkmann, Florian, Joudaki, Amir, Boeva, Valentina2025